1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/ext2_obd/ext2_obd.c
6 * Copyright (C) 2001 Cluster File Systems, Inc.
8 * This code is issued under the GNU General Public License.
9 * See the file COPYING in this distribution
11 * by Peter Braam <braam@clusterfs.com>
16 #include <linux/version.h>
17 #include <linux/module.h>
19 #include <linux/stat.h>
20 #include <linux/locks.h>
21 #include <linux/ext2_fs.h>
22 #include <linux/quotaops.h>
23 #include <asm/unistd.h>
25 #define DEBUG_SUBSYSTEM S_FILTER
27 #include <linux/obd_class.h>
28 #include <linux/obd_ext2.h>
29 #include <linux/obd_filter.h>
31 extern struct obd_device obd_dev[MAX_OBD_DEVICES];
34 #define FILTER_ROOTINO 2
37 static char * obd_type_by_mode[S_IFMT >> S_SHIFT] = {
39 [S_IFREG >> S_SHIFT] "R",
40 [S_IFDIR >> S_SHIFT] "D",
41 [S_IFCHR >> S_SHIFT] "C",
42 [S_IFBLK >> S_SHIFT] "B",
43 [S_IFIFO >> S_SHIFT] "F",
44 [S_IFSOCK >> S_SHIFT] "S",
45 [S_IFLNK >> S_SHIFT] "L"
49 /* write the pathname into the string */
50 static void filter_id(char *buf, obd_id id, obd_mode mode)
52 sprintf(buf, "O/%s/%Ld", obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT],
56 /* setup the object store with correct subdirectories */
57 static int filter_prep(struct obd_device *obddev)
59 struct obd_run_ctxt saved;
67 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
68 rc = simple_mkdir(current->fs->pwd, "O", 0700);
69 rc = simple_mkdir(current->fs->pwd, "P", 0700);
70 rc = simple_mkdir(current->fs->pwd, "D", 0700);
71 file = filp_open("O", O_RDONLY, 0);
73 CERROR("cannot open O\n");
74 GOTO(out, rc = PTR_ERR(file));
76 rc = simple_mkdir(file->f_dentry, "R", 0700); /* regular */
77 rc = simple_mkdir(file->f_dentry, "D", 0700); /* directory */
78 rc = simple_mkdir(file->f_dentry, "L", 0700); /* symbolic links */
79 rc = simple_mkdir(file->f_dentry, "C", 0700); /* character devices */
80 rc = simple_mkdir(file->f_dentry, "B", 0700); /* block devices */
81 rc = simple_mkdir(file->f_dentry, "F", 0700); /* fifo's */
82 rc = simple_mkdir(file->f_dentry, "S", 0700); /* sockets */
83 filp_close(file, NULL);
85 filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
86 file = filp_open(rootid, O_RDWR | O_CREAT, 00755);
88 CERROR("OBD filter: cannot make root directory");
89 GOTO(out, rc = PTR_ERR(file));
92 /* FIXME: this is the same as the _file_ we just created? */
93 rc = simple_mkdir(current->fs->pwd, rootid, 0755);
95 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
96 if ( !file || IS_ERR(file) ) {
97 CERROR("OBD filter: cannot open/create status file\n");
98 GOTO(out, rc = PTR_ERR(file));
101 /* steal operations */
102 inode = file->f_dentry->d_inode;
103 obddev->u.filter.fo_fop = file->f_op;
104 obddev->u.filter.fo_iop = inode->i_op;
105 obddev->u.filter.fo_aops = inode->i_mapping->a_ops;
108 if (inode->i_size == 0) {
109 ssize_t retval = file->f_op->write(file, (char *)&lastino,
110 sizeof(lastino), &off);
111 if (retval != sizeof(lastino)) {
112 CERROR("OBD filter: error writing lastino\n");
113 GOTO(out, rc = -EIO);
116 ssize_t retval = file->f_op->read(file, (char *)&lastino,
117 sizeof(lastino), &off);
118 if (retval != sizeof(lastino)) {
119 CERROR("OBD filter: error reading lastino\n");
120 GOTO(out, rc = -EIO);
123 obddev->u.filter.fo_lastino = lastino;
133 /* cleanup the filter: write last used object id to status file */
134 static void filter_post(struct obd_device *obddev)
136 struct obd_run_ctxt saved;
141 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
142 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
143 if ( !file || IS_ERR(file)) {
144 CERROR("OBD filter: cannot create status file\n");
147 rc = file->f_op->write(file, (char *)&obddev->u.filter.fo_lastino,
148 sizeof(obddev->u.filter.fo_lastino), &off);
149 if (rc != sizeof(obddev->u.filter.fo_lastino) ) {
150 CERROR("OBD filter: error writing lastino\n");
153 rc = filp_close(file, NULL);
155 CERROR("OBD filter: cannot close status file\n");
162 static __u64 filter_next_id(struct obd_device *obddev)
165 spin_lock(&obddev->u.filter.fo_lock);
166 obddev->u.filter.fo_lastino++;
167 id = obddev->u.filter.fo_lastino;
168 spin_unlock(&obddev->u.filter.fo_lock);
172 /* how to get files, dentries, inodes from object id's */
173 static struct file *filter_obj_open(struct obd_device *obddev,
174 __u64 id, __u32 type)
176 struct obd_run_ctxt saved;
178 struct super_block *sb;
181 sb = obddev->u.filter.fo_sb;
182 if (!sb || !sb->s_dev) {
183 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
189 CDEBUG(D_INODE, "fatal: invalid obdo %Lu\n", id);
194 if ( ! (type & S_IFMT) ) {
195 CERROR("OBD filter_obj_open, no type (%Ld), mode %o!\n",
199 filter_id(name, id, type);
200 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
201 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0);
204 CDEBUG(D_INODE, "opening obdo %s\n", name);
209 static struct file *filter_parent(obd_id id, obd_mode mode)
214 sprintf(path, "O/%s", obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT]);
216 file = filp_open(path, O_RDONLY, 0);
221 static struct inode *filter_inode_from_obj(struct obd_device *obddev,
222 __u64 id, __u32 type)
227 file = filter_obj_open(obddev, id, type);
229 CERROR("filter_inode_from_obdo failed\n");
233 inode = iget(file->f_dentry->d_inode->i_sb,
234 file->f_dentry->d_inode->i_ino);
240 static int filter_connect(struct obd_conn *conn)
245 rc = gen_connect(conn);
253 static int filter_disconnect(struct obd_conn *conn)
257 rc = gen_disconnect(conn);
261 /* XXX cleanup preallocated inodes */
265 /* mount the file system (secretly) */
266 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
268 struct obd_ioctl_data* data = buf;
269 struct vfsmount *mnt;
273 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
277 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
282 /* XXX is this even possible if do_kern_mount succeeded? */
283 obddev->u.filter.fo_sb = mnt->mnt_root->d_inode->i_sb;
284 if (!obddev->u.filter.fo_sb)
285 GOTO(err_put, err = -ENODEV);
287 obddev->u.filter.fo_vfsmnt = mnt;
288 obddev->u.filter.fo_fstype = strdup(data->ioc_inlbuf2);
290 obddev->u.filter.fo_ctxt.pwdmnt = mnt;
291 obddev->u.filter.fo_ctxt.pwd = mnt->mnt_root;
292 obddev->u.filter.fo_ctxt.fs = KERNEL_DS;
294 err = filter_prep(obddev);
296 GOTO(err_kfree, err);
297 spin_lock_init(&obddev->u.filter.fo_lock);
302 kfree(obddev->u.filter.fo_fstype);
305 mntput(obddev->u.filter.fo_vfsmnt);
306 obddev->u.filter.fo_sb = 0;
315 static int filter_cleanup(struct obd_device * obddev)
317 struct super_block *sb;
321 if ( !(obddev->obd_flags & OBD_SET_UP) )
324 if ( !list_empty(&obddev->obd_gen_clients) ) {
325 CERROR("still has clients!\n");
329 sb = obddev->u.filter.fo_sb;
330 if (!obddev->u.filter.fo_sb)
336 mntput(obddev->u.filter.fo_vfsmnt);
337 obddev->u.filter.fo_sb = 0;
338 kfree(obddev->u.filter.fo_fstype);
347 static inline void filter_from_inode(struct obdo *oa, struct inode *inode)
349 int type = oa->o_mode & S_IFMT;
352 CDEBUG(D_INFO, "src inode %ld, dst obdo %ld valid 0x%08x\n",
353 inode->i_ino, (long)oa->o_id, oa->o_valid);
354 obdo_from_inode(oa, inode);
355 oa->o_mode &= ~S_IFMT;
358 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
359 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
360 CDEBUG(D_INODE, "copying device %x from inode to obdo\n",
362 *((obd_rdev *)oa->o_inline) = rdev;
363 oa->o_obdflags |= OBD_FL_INLINEDATA;
364 oa->o_valid |= OBD_MD_FLINLINE;
368 else if (filter_has_inline(inode)) {
369 CDEBUG(D_INFO, "copying inline from inode to obdo\n");
370 memcpy(oa->o_inline, inode->u.ext2_i.i_data,
371 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
372 oa->o_obdflags |= OBD_FL_INLINEDATA;
373 oa->o_valid |= OBD_MD_FLINLINE;
376 if (filter_has_obdmd(inode)) {
377 /* XXX this will change when we don't store the obdmd in data */
378 CDEBUG(D_INFO, "copying obdmd from inode to obdo\n");
379 memcpy(oa->o_obdmd, inode->u.ext2_i.i_data,
380 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
381 oa->o_obdflags |= OBD_FL_OBDMDEXISTS;
382 oa->o_valid |= OBD_MD_FLOBDMD;
388 static int filter_getattr(struct obd_conn *conn, struct obdo *oa)
393 if ( !gen_client(conn) ) {
394 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
399 if ( !(inode = filter_inode_from_obj(conn->oc_dev,
400 oa->o_id, oa->o_mode)) ) {
405 oa->o_valid &= ~OBD_MD_FLID;
406 filter_from_inode(oa, inode);
413 static int filter_setattr(struct obd_conn *conn, struct obdo *oa)
420 if (!gen_client(conn)) {
421 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
425 inode = filter_inode_from_obj(conn->oc_dev, oa->o_id, oa->o_mode);
431 iattr_from_obdo(&iattr, oa);
432 iattr.ia_mode &= ~S_IFMT;
433 iattr.ia_mode |= S_IFREG;
435 if ( inode->i_op->setattr ) {
436 rc = inode->i_op->setattr(&de, &iattr);
438 rc = inode_setattr(inode, &iattr);
446 static int filter_open(struct obd_conn *conn, struct obdo *oa)
451 if (!gen_client(conn))
454 if ( !(inode = filter_inode_from_obj(conn->oc_dev,
455 oa->o_id, oa->o_mode)) )
461 static int filter_close(struct obd_conn *conn, struct obdo *oa)
466 if (!gen_client(conn))
469 if ( !(inode = filter_inode_from_obj(conn->oc_dev,
470 oa->o_id, oa->o_mode)) )
473 iput(inode); /* for the close */
474 iput(inode); /* for this call */
478 static int filter_create (struct obd_conn* conn, struct obdo *oa)
481 struct obd_run_ctxt saved;
484 struct obd_device *obddev = conn->oc_dev;
488 if (!gen_client(conn)) {
489 CERROR("invalid client %u\n", conn->oc_id);
493 oa->o_id = filter_next_id(conn->oc_dev);
494 if ( !(oa->o_mode && S_IFMT) ) {
495 CERROR("filter obd: no type!\n");
499 filter_id(name, oa->o_id, oa->o_mode);
500 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
504 file = filp_open(name, O_RDONLY | O_CREAT, mode);
507 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
512 /* Set flags for fields we have set in ext2_new_inode */
513 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
514 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME |
515 OBD_MD_FLUID | OBD_MD_FLGID;
519 static int filter_destroy(struct obd_conn *conn, struct obdo *oa)
521 struct obd_device * obddev;
522 struct obd_client * cli;
523 struct inode * inode;
527 struct obd_run_ctxt saved;
529 if (!(cli = gen_client(conn))) {
530 CERROR("invalid client %u\n", conn->oc_id);
535 obddev = conn->oc_dev;
536 object = filter_obj_open(obddev, oa->o_id, oa->o_mode);
537 if (!object || IS_ERR(object)) {
542 inode = object->f_dentry->d_inode;
544 inode->i_mode = 010000;
546 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
547 dir = filter_parent(oa->o_id, oa->o_mode);
554 dget(object->f_dentry);
555 rc = vfs_unlink(dir->f_dentry->d_inode, object->f_dentry);
558 filp_close(object, 0);
565 static int filter_truncate(struct obd_conn *conn, struct obdo *oa, obd_size count,
570 error = filter_setattr(conn, oa);
571 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
577 /* buffer must lie in user memory here */
578 static int filter_read(struct obd_conn *conn, struct obdo *oa, char *buf,
579 obd_size *count, obd_off offset)
582 unsigned long retval;
585 if (!gen_client(conn)) {
586 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
591 file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
592 if (!file || IS_ERR(file)) {
594 return -PTR_ERR(file);
597 /* count doubles as retval */
598 retval = file->f_op->read(file, buf, *count, &offset);
613 /* buffer must lie in user memory here */
614 static int filter_write(struct obd_conn *conn, struct obdo *oa, char *buf,
615 obd_size *count, obd_off offset)
619 unsigned long retval;
622 if (!gen_client(conn)) {
623 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
628 file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
629 if (!file || IS_ERR(file)) {
631 return -PTR_ERR(file);
634 /* count doubles as retval */
635 retval = file->f_op->write(file, buf, *count, &offset);
651 static int filter_pgcache_brw(int rw, struct obd_conn *conn,
660 struct super_block *sb;
662 int onum; /* index to oas */
663 int pnum; /* index to pages (bufs) */
664 unsigned long retval;
670 if (!gen_client(conn)) {
671 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
676 sb = conn->oc_dev->u.filter.fo_sb;
680 pnum = 0; /* pnum indexes buf 0..num_pages */
681 for (onum = 0; onum < num_oa; onum++) {
684 file = filter_obj_open(conn->oc_dev, oa[onum]->o_id,
686 if (!file || IS_ERR(file)) {
692 /* count doubles as retval */
693 for (pg = 0; pg < oa_bufs[onum]; pg++) {
694 CDEBUG(D_INODE, "OP %d obdo no/pno: (%d,%d) (%ld,%ld) off count (%Ld,%Ld)\n",
695 rw, onum, pnum, file->f_dentry->d_inode->i_ino,
696 (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
697 offset[pnum], count[pnum]);
702 buffer = kmap(pages[pnum]);
703 retval = file->f_op->write(file, buffer, count[pnum], &off);
705 CDEBUG(D_INODE, "retval %ld\n", retval);
707 loff_t off = offset[pnum];
708 char *buffer = kmap(pages[pnum]);
710 if (off >= file->f_dentry->d_inode->i_size) {
711 memset(buffer, 0, count[pnum]);
712 retval = count[pnum];
714 retval = file->f_op->read(file, buffer, count[pnum], &off);
718 if ( retval != count[pnum] ) {
724 CDEBUG(D_INODE, "retval %ld\n", retval);
728 /* sizes and blocks are set by generic_file_write */
729 /* ctimes/mtimes will follow with a setattr call */
736 error = (retval >= 0) ? 0 : retval;
741 struct inode *ioobj_to_inode(struct obd_conn *conn, struct obd_ioobj *o)
743 struct inode *inode = NULL;
744 struct super_block *sb = conn->oc_dev->u.ext2.e2_sb;
746 if (!sb || !sb->s_dev) {
747 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
753 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)o->ioo_id);
758 inode = filter_inode_from_obj(conn->oc_dev, o->ioo_id, S_IFREG);
759 if (!inode || inode->i_nlink == 0 || is_bad_inode(inode)) {
760 CERROR("from obdo - fatal: invalid inode %ld (%s).\n",
761 (long)o->ioo_id, inode ? inode->i_nlink ? "bad inode" :
762 "no links" : "NULL");
772 static int filter_preprw(int cmd, struct obd_conn *conn,
773 int objcount, struct obd_ioobj *obj,
774 int niocount, struct niobuf *nb,
777 struct obd_ioobj *o = obj;
778 struct niobuf *b = nb;
779 struct niobuf *r = res;
783 memset(res, 0, sizeof(*res) * niocount);
785 for (i = 0; i < objcount; i++, o++) {
787 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
788 unsigned long index = b->offset >> PAGE_SHIFT;
789 struct inode *inode = ioobj_to_inode(conn, o);
792 /* FIXME: we need to iput all inodes on error */
796 if (cmd == OBD_BRW_WRITE)
797 page = lustre_get_page_write(inode, index);
799 page = lustre_get_page_read(inode, index);
801 RETURN(PTR_ERR(page));
803 r->addr = (__u64)(unsigned long)page_address(page);
804 r->offset = b->offset;
812 static int filter_commitrw(int cmd, struct obd_conn *conn,
813 int objcount, struct obd_ioobj *obj,
814 int niocount, struct niobuf *res)
816 struct obd_ioobj *o = obj;
817 struct niobuf *r = res;
821 for (i = 0; i < objcount; i++, obj++) {
823 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
824 struct page *page = r->page;
829 if (cmd == OBD_BRW_WRITE) {
830 int rc = lustre_commit_page(page, 0, PAGE_SIZE);
832 /* FIXME: still need to iput the other inodes */
836 lustre_put_page(page);
838 iput(page->mapping->host);
844 static int filter_statfs (struct obd_conn *conn, struct statfs * statfs)
846 struct super_block *sb;
851 if (!gen_client(conn)) {
852 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
857 sb = conn->oc_dev->u.filter.fo_sb;
859 err = sb->s_op->statfs(sb, statfs);
862 } /* filter_statfs */
865 static int filter_get_info(struct obd_conn *conn, obd_count keylen,
866 void *key, obd_count *vallen, void **val)
868 struct obd_device *obddev;
869 struct obd_client * cli;
872 if (!(cli = gen_client(conn))) {
873 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
877 obddev = conn->oc_dev;
879 if ( keylen == strlen("blocksize") &&
880 memcmp(key, "blocksize", keylen) == 0 ) {
881 *vallen = sizeof(int);
882 *val = (void *)obddev->u.filter.fo_sb->s_blocksize;
887 if ( keylen == strlen("blocksize_bits") &&
888 memcmp(key, "blocksize_bits", keylen) == 0 ){
889 *vallen = sizeof(int);
890 *val = (void *)(int)obddev->u.filter.fo_sb->s_blocksize_bits;
895 if ( keylen == strlen("root_ino") &&
896 memcmp(key, "root_ino", keylen) == 0 ){
897 *vallen = sizeof(int);
898 *val = (void *)(int) FILTER_ROOTINO;
903 CDEBUG(D_IOCTL, "invalid key\n");
908 struct obd_ops filter_obd_ops = {
910 o_get_info: filter_get_info,
911 o_setup: filter_setup,
912 o_cleanup: filter_cleanup,
913 o_connect: filter_connect,
914 o_disconnect: filter_disconnect,
915 o_statfs: filter_statfs,
916 o_getattr: filter_getattr,
917 o_create: filter_create,
918 o_setattr: filter_setattr,
919 o_destroy: filter_destroy,
921 o_close: filter_close,
923 o_write: filter_write,
924 o_brw: filter_pgcache_brw,
925 o_punch: filter_truncate,
926 o_preprw: filter_preprw,
927 o_commitrw: filter_commitrw
929 o_preallocate: filter_preallocate_inodes,
930 o_migrate: filter_migrate,
931 o_copy: gen_copy_data,
932 o_iterate: filter_iterate
937 static int __init obdfilter_init(void)
939 printk(KERN_INFO "Filtering OBD driver v0.001, braam@clusterfs.com\n");
940 return obd_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
943 static void __exit obdfilter_exit(void)
945 obd_unregister_type(OBD_FILTER_DEVICENAME);
948 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
949 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
950 MODULE_LICENSE("GPL");
952 module_init(obdfilter_init);
953 module_exit(obdfilter_exit);